Case Study
Automated Product Tagging
for a C2C Online Marketplace
Company |
Industry |
 |
E-Commerce |
Client
Willhaben is Austria's largest free online marketplace, with more than 10 million currently listed items. Each day, approximately 400,000 ads are posted on the platform. The marketplace covers various categories, including real estate, automobiles and motorcycles, job listings, and a miscellaneous section aptly named \"Marketplace.\" Notably, the fashion category, which includes clothing, shoes, and accessories, is one of the largest segments within the miscellaneous categories. The Willhaben website ranks among Austria's most visited websites, attracting over 8 million unique users monthly.
Impact
The online marketplace, Willhaben, has significantly reduced the time required for users to post ads by integrating our intelligent AI suggestions for attributes like color, pattern, and brand. This has not only encouraged users to upload more items but has also increased their engagement. With the introduction of automatic attribute suggestions, users are more motivated to provide accurate and detailed information about the product they are selling. This, in turn, has elevated the overall content quality.
Problem
When operating a large online marketplace, it's crucial to ensure that users can easily find the items they intend to purchase. A tremendously helpful tool for achieving this goal is the implementation of filters on the website or mobile app, particularly for valuable second-hand online marketplaces where any user generates content and products. For fashion-related items, these filters include various attributes, such as size, location, color, pattern, brand, and more.
Ensuring all items possess these attributes as required fields is essential to render the filters effective. However, the process of selecting and inputting all these attributes during the posting of new items can be both tedious and time-consuming. This project aimed to develop an ML model capable of suggesting suitable colors, patterns and brands based on the images and titles provided by the users to reduce the effort and time required to post an ad.
Challenges
Model quality and speed optimization
The main challenge was creating a high-quality model that boasts optimal speed and is fast enough.
Inaccurate dataset
There is a dataset of user ads on Willhaben, but the user labels associated with these ads frequently contain inaccuracies.
Scalability and efficient processing
Furthermore, developing a model that can swiftly process substantial daily request volumes while maintaining low latency levels was of high priority in ensuring a seamless user experience.
Solution
We trained a vision model based on the transformer architecture, specifically focusing on colors and patterns. To enhance the accuracy of color predictions, we used labels from our previously developed model, which had been trained on a manually labeled color dataset. This model has an additional output head designed for predicting patterns. In the case of patterns, our approach was to include user annotations from Willhaben ads.
To complement our vision model, we also implemented a text-based brand detection model. This model employs the content of titles to forecast brands and can predict from a selection of over 100 different brands. To evaluate the performance of both models, we tested them using a manually labeled set of items.
Tools and Technologies
Python
PyTorch
Weights And Biases
Falcon
Triton Inference Server
Google Cloud Platform
Results
The models were then deployed in the cloud and used by Willhaben to speed up their ad posting process and improve the quality of item filters. This improvement in efficient filtering and enhanced search functionality has increased user experience as they can now effortlessly find what they're looking for, fostering a more personalized and valuable shopping experience.
Discover More Case Studies
Discover the impact of our custom AI solutions on business success through customer stories
